Edge2Prompt: Modality-Agnostic Model for Out-of-Distribution Liver Segmentation

Nathan Hollet; Oumeymah Cherkaoui; Philippe C. Cattin; Sidaty El Hadramy

REVIEW 3 major objections 5 minor 29 references

Edge2Prompt: Modality-Agnostic Model for Out-of-Distribution Liver Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Edge2Prompt claims a single pipeline segments liver across CT and MRI by prompting SAM-2 with logits derived from edge maps.

desk verdict Reasonable idea, but the headline OOD gain is against the weakest baseline and the data-scarce test set doesn't add up; deserves review after fixing the protocol. read the letter →

arxiv 2508.04305 v3 pith:ORGTEXZO submitted 2025-08-06 eess.IV

Nathan Hollet , Oumeymah Cherkaoui , Philippe C. Cattin , Sidaty El Hadramy This is my paper

classification eess.IV

keywords liversegmentationmodality-agnosticmodelout-of-distributiongeneralizationSAM-2edgedetectionpromptgenerationCHAOSdatasetdata-efficient

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to make liver segmentation from abdominal scans work when the scanner modality changes, such as training on MRI and testing on CT. It proposes Edge2Prompt, a pipeline that turns each 2D slice into a binarized edge map, runs that map through a small U-Net to produce logit prompts, and feeds those prompts to a frozen SAM-2 model to generate the final mask. On the CHAOS dataset, the method reaches 86.4% mean Dice on out-of-distribution tasks, beating U-Net baselines by 27.4 percentage points and another self-prompting SAM-2 method by 9.1 points. The point is that a deliberately simple, classical preprocessing step can make a large foundation model usable in low-data, cross-modality clinical settings.

What carries the argument

Edge-to-prompt transfer: histogram-equalized Canny edge maps are fed into a modified U-Net whose output logits serve as the prompt for a frozen SAM-2, connected through a differentiable wrapper so gradients flow back to the U-Net. The logit prompt replaces manual point- or box-based prompting, the binarized edges strip away intensity-based modality cues, and the frozen SAM-2 supplies the shape prior and refinement.

What would settle it

Train Edge2Prompt on MR T1-DUAL only. On CT test slices, corrupt the edge maps by randomly dropping half the boundary pixels or flipping edge/non-edge labels. If out-of-distribution Dice stays near 85%, the edge-to-prompt mapping is not what carries the result; if it collapses toward the U-Net baseline of about 62%, edge-map fidelity is the mechanism.

Watch

Extended reading notes

Core claim

The central claim is that binarized Canny edge maps, extracted after histogram equalization, preserve enough liver-boundary structure across CT and MRI that a single U-Net can learn transferable logit prompts for SAM-2. Edge2Prompt is trained end-to-end while all SAM-2 parameters stay frozen; the composite loss of focal, dice, and log-cosh dice terms makes the prompt-generating U-Net learn a task that SAM-2 can refine. With training on a single CT and a single MR volume, or on one modality alone, the pipeline reconstructs coherent 3D liver volumes from segmented slices and outperforms standard U-Nets and a coarse-mask self-prompting baseline on out-of-distribution data.

Load-bearing premise

The whole out-of-distribution gain rests on one premise: that binarized Canny edges preserve the same liver boundary structure in CT and MRI, so a U-Net trained on one modality's edges can prompt SAM-2 for the other.

Editorial extensions

If this is right

One trained model can segment livers in both CT and MRI without retraining, because edge maps remove the intensity and contrast cues that differ between modalities.
Training on a single CT and a single MR volume is enough for usable out-of-distribution segmentation, which directly addresses annotation scarcity in clinical deployment.
SAM-2 stays frozen during training, so the pipeline does not need to fine-tune a large foundation model and remains cheap to deploy and update.
Segmented 2D slices can be stacked into 3D volumes via an affine transform, giving clinicians reconstruction-ready output for preoperative planning.
The method outperforms U-Net baselines and another self-prompting SAM-2 approach on cross-modality generalization, suggesting the edge-prompt path is more robust than raw-image or coarse-mask prompting.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The ablation showing Sobel edges nearly matching Canny suggests the exact edge detector is not the load-bearing choice; what matters may be that the edge map is binarized and topology-preserving. A simpler binary boundary operator could work equally well.
If edge maps are truly the modality-agnostic carrier, the same prompt-generating U-Net should transfer to other organs or to ultrasound and fluoroscopy with only small retraining; the paper does not test this, but it is a direct consequence of the stated mechanism.
Because the U-Net output is a logit prompt rather than a final mask, the model may be learning a task-specific prior that SAM-2 refines; this predicts that failures will concentrate on small or faint liver regions, which is consistent with the high variance the paper reports.
A testable extension: corrupting edge maps at test time (randomly dropping or flipping boundary pixels) should degrade out-of-distribution Dice much more than in-distribution Dice, since cross-modality transfer depends on boundary consistency.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: this is a sensible engineering contribution with a clear pipeline—Canny edge maps, a U-Net prompt generator, and frozen SAM-2—and the OOD results are still positive against the stronger edge-only baseline. But the abstract's 27.4% gain is cherry-picked from imU-Net, and the data-scarce test set is not derivable from the stated split, which makes Table 1 unverifiable.

What's actually new: combining edge maps with a U-Net to produce logit prompts for SAM-2 is different from earlier self-prompting via logistic regression. The paper isolates components well (imU-Net, emU-Net, sU-Net, spSAM, Ours) and the ablations are honest—including the finding that Sobel edges perform nearly as well as Canny, which quietly undercuts the claim that edge-map choice is critical. The loss ablation is also reasonable, and the cross-modality setup (train CT, test MR and vice versa) is a fair OOD test.

Soft spots, in proportion: The stress-test is correct. In §3.2, you split into 20 train / 10 val / 10 test, then introduce D_TRAIN_J (1 CT + 1 MR) and D_TEST_J (14 CT + 14 MR). You can't form a 14+14 test set from the remaining 19+19 without either borrowing from validation/test or re-splitting, and the text never says which. This puts all Table 1 data-scarce comparisons in question. The 27.4% figure is only versus imU-Net; against emU-Net, the average OOD gain is about 8.8 Dice points. Still meaningful, but not the headline number. Also, the paper never directly tests whether edge maps are truly modality-agnostic; the Sobel result suggests the exact edge detector is not the source of transferability, so the 'modality-agnostic edge map' story is weaker than claimed. Code is promised but not available, so details like Canny thresholds and how D_TEST_J is constructed aren't checkable.

The paper is coherent on its own terms, not circular, and the authors report results that partially undercut their own narrative—that's a sign of good faith. It's not a field-reshaping result, but it's a useful practical contribution for SAM-based medical segmentation. If the split issue is clarified and the abstract is reframed around the stronger baseline, it could be a solid conference paper. As is, it deserves peer review but with the protocol questions as mandatory revisions.

Recommendation: send to peer review, but ask the authors to explain exactly how D_TEST_J was formed and to report margins against emU-Net in the abstract.

Referee Report

3 major / 5 minor

Summary. The paper proposes Edge2Prompt, a segmentation pipeline that combines classical edge detection with a trainable U-Net and a frozen SAM-2 model. Given a 2D grayscale CT or MR slice, an Edge Detection Module extracts a Canny edge map, a U-Net converts that edge map into logit-based prompts, and SAM-2 uses those prompts to produce a refined liver segmentation. Segmented slices are then stacked into a 3D volume. The method is evaluated on the CHAOS dataset under three scenarios: in-distribution mixed-modality training, a data-scarce setting with two training volumes, and cross-modality out-of-distribution generalization (train on CT, test on MR and vice versa). The authors report a mean OOD Dice of 86.4%, claiming a 27.4% improvement over U-Net baselines and a 9.1% improvement over another self-prompting SAM-2 method.

Significance. If the reported results are reproducible, the paper makes a useful empirical contribution: it demonstrates that a lightweight prompt generator trained on cheap edge maps can condition a frozen foundation model to segment the liver across CT and MRI with little training data. The end-to-end differentiable wrapper around frozen SAM-2, the composite loss ablation, and the systematic comparison against imU-Net, emU-Net, sU-Net, and spSAM are strengths. The method does not require fine-tuning SAM-2, which is practical for clinical settings with limited compute. However, the main quantitative claims rest on an under-specified data split and a selective baseline comparison. The data-scarce results in Table 1 are not verifiable from the described partition, and the abstract's headline margin is computed against only the weakest U-Net baseline. These issues are fixable but must be addressed before the central claims can be accepted.

major comments (3)

[§3.2 / Table 1] The evaluation protocol is numerically inconsistent. The authors state D_TRAIN_T = 20 volumes (10 CT + 10 MR), D_VAL_T = 10, and D_TEST_T = 10, which accounts for all 40 CHAOS volumes. They then introduce D_TRAIN_J = 1 CT + 1 MR and D_TEST_J = 14 CT + 14 MR. These 28 test volumes cannot be formed from the stated partition: after removing D_TRAIN_J, only 19 CT + 19 MR volumes remain, and the original validation/test sets contain only 10 CT + 10 MR combined. If D_TEST_J includes volumes from D_TRAIN_T or D_VAL_T, then the Table 1 rows are not comparable and the data-scarce results may reflect a different, potentially easier, test distribution. Please disclose the exact volume-level split, including patient/scan identifiers, and rerun or re-report the data-scarce evaluation on a held-out set that does not overlap with the volumes used elsewhere.
[Abstract / §4.1 / Table 2] The claim of 'outperforming U-Net baselines by 27.4%' is an average over only imU-Net, the weakest baseline. From Table 2, the OOD Dice differences are CT: 87.6−62.2 = 25.4 and MR: 85.3−55.9 = 29.4, whose mean is 27.4. Against emU-Net, the U-Net baseline that also uses edge maps, the average improvement is (87.6−80.2 + 85.3−75.2)/2 = 8.8 percentage points. Since the abstract refers to 'U-Net baselines' in the plural, the current wording materially overstates the method's advantage over the stronger edge-based baseline. Please report margins against each baseline separately and adjust the abstract.
[§2.1 / Table 3b] The central methodological premise is that Canny edge maps after histogram equalization are 'modality-agnostic' and preserve transferable liver-boundary structure across CT and MRI. This invariance is never directly tested. The ablation in Table 3b shows Sobel and Canny produce nearly identical OOD Dice (85.2 vs 85.3), which suggests the specific edge detector is not the source of robustness; it does not establish that edge maps are modality-agnostic. I recommend adding a direct analysis, e.g., a qualitative/quantitative comparison of edge maps from co-registered or corresponding CT/MR slices, or an ablation that replaces edge-map input with raw intensity input in the same SAM-2-conditioned architecture (the comparison with sU-Net is suggestive but sU-Net also differs in U-Net input). Without this, the mechanistic explanation for the OOD gain remains a conjecture.

minor comments (5)

[§3.2] 'T1-DUAL OOP' is ambiguous. Earlier the paper mentions 'T1-DUAL (In-Phase/Out-Phase)'; please define 'OOP' explicitly (presumably out-of-phase).
[§4.1] The claim of '+14.4% mDICE' over spSAM on ID tasks does not match the numbers in Tables 1 and 2. Table 1 gives 95.1−78.8 = 16.3; Table 2 gives an average of (95.1−77.8 + 92.8−81.1)/2 = 14.5. Please clarify how 14.4% was computed.
[§2.3] The volumetric reconstruction section mentions interpolation or super-resolution for sparse MR slices, but no quantitative evaluation of the reconstruction quality is provided beyond two visual cases. This is acceptable as a secondary contribution, but a sentence stating that the 3D reconstruction is illustrative rather than quantitatively validated would help.
[General] The source code is only announced as 'will be made publicly available.' For reproducibility, provide a code link or an artifact in the submission.
[Table 3b] The Laplacian detector has a very high standard deviation (69.6 ± 37.2 Dice). It would be informative to report the number of failing cases (e.g., near-zero Dice) to help interpret the variance.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; Edge2Prompt's reported segmentations are genuine model outputs, and the data-split / model-selection concerns are validity issues, not derivation equivalences.

full rationale

Edge2Prompt is an empirical machine-learning pipeline, not a derivation from first principles. The only 'prediction' is the 2D/3D liver segmentation, produced by a frozen SAM-2 conditioned on U-Net logits, and evaluated on the CHAOS dataset against external and internal baselines. I could not find any equation, definition, or fitting operation that makes the reported Dice scores equal to an input or to a self-citation by construction. Loss weights (wF=2, wD=2, wLCD=3) and the Canny edge detector are chosen empirically, but they are hyperparameters rather than parameters fitted to the test labels; the reported segmentation is still an independent function of the input image. The ablation in Table 3a selects the best loss configuration on the OOD test task, which is an optimistic model-selection practice, but it does not make the OOD result an identity. The only author self-citation (Ref. [11], by co-author S. El hadramy) appears as background in the introduction and is not load-bearing for any technical claim. I also note a reproducible-validity issue outside circularity: the §3.2 partition is internally inconsistent—D_TRAIN_T (10 CT + 10 MR), D_VAL_T (5 CT + 5 MR), and D_TEST_T (5 CT + 5 MR) account for all 40 volumes, so a later D_TEST_J of 14 CT + 14 MR after removing a 1+1 training subset cannot be derived from the stated split. This undermines the data-scarce benchmark clarity but is not a circularity of the method's derivation chain. Overall, no circular step was found.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; all components (Canny, U-Net, SAM-2) are pre-existing. The free parameters are edge detection settings and loss weights, both chosen by hand or validation.

free parameters (2)

Canny edge detector thresholds and sigma = not specified in paper
The Kornia Canny filter has low/high thresholds and a Gaussian sigma that are not reported; they directly control which edges become prompts and could affect OOD transfer.
Composite loss weights wF, wD, wLCD = wF=2, wD=2, wLCD=3
Stated as 'empirically chosen' in Section 3.1. These weights balance focal, dice, and log-cosh dice losses and affect how the U-Net learns prompts.

assumptions (3)

domain assumption Canny edges from histogram-equalized images are modality-agnostic across CT and MR
The core premise of EDM (Section 2.1). If edges do not preserve the same liver boundary structure across modalities, the U-Net cannot learn transferable prompts.
domain assumption SAM-2 can refine liver masks from U-Net logit prompts without fine-tuning
Section 2.2 freezes all SAM-2 parameters and relies on its pretrained representations to interpret logits as prompts; if SAM-2 cannot work with logit prompts on medical images, the pipeline fails.
domain assumption CHAOS dataset is representative of clinical CT/MR liver imaging
All evaluation is on CHAOS. External validity to other scanners, protocols, or hospitals is not tested, so the 'modality-agnostic' claim is limited to this dataset.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Edge2Prompt: Modality-Agnostic Model for Out-of-Distribution Liver Segmentation." pith.science (2026). https://pith.science/paper/ORGTEXZO

@misc{pith2026250804305,
  author       = {Pith},
  title        = {Pith review of: Edge2Prompt: Modality-Agnostic Model for Out-of-Distribution Liver Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ORGTEXZO}},
  note         = {Machine review of arXiv:2508.04305}
}

read the original abstract

Liver segmentation is essential for preoperative planning in interventions like tumor resection or transplantation, but implementation in clinical workflows faces challenges due to modality-specific tools and data scarcity. We propose Edge2Prompt, a novel pipeline for modality-agnostic liver segmentation that generalizes to out-of-distribution (OOD) data. Our method integrates classical edge detection with foundation models. Modality-agnostic edge maps are first extracted from input images, then processed by a U-Net to generate logit-based prompts. These prompts condition the Segment Anything Model 2 (SAM-2) to generate 2D liver segmentations, which can then be reconstructed into 3D volumes. Evaluated on the multi-modal CHAOS dataset, Edge2Prompt achieves competitive results compared to classical segmentation methods when trained and tested in-distribution (ID), and outperforms them in data-scarce scenarios due to the SAM-2 module. Furthermore, it achieves a mean Dice Score of 86.4% on OOD tasks, outperforming U-Net baselines by 27.4% and other self-prompting methods by 9.1%, demonstrating its effectiveness. This work bridges classical and foundation models for clinically adaptable, data-efficient segmentation.

Figures

Figures reproduced from arXiv: 2508.04305 by the authors.

**Figure 1.** General workflow: The Edge Detection Module (EDM) extracts modality-agnostic edge maps from input images, which the Edge2Prompt block processes to generate refined segmentation mask predictions, enabling 3D organ model reconstruction from segmented slices. 2.1 From 2D CT/MR Images to Edge Maps The Edge Detection Module (EDM) first preprocesses the input grayscale image through histogram equalization and bilinear fi… view at source ↗

**Figure 2.** Edge2Prompt training workflow: Modality-agnostic edge maps generated by the EDM serve as input to a modified U-Net architecture. This U-Net processes the edge information to generate a logit-based prompt that conditions a frozen SAM-2 module. The prompted SAM-2 module then generates refined segmentation mask predictions. Using a composite loss function, these predictions are then evaluated against ground-truth mas… view at source ↗

**Figure 3.** Qualitative comparison of 3D CT volume reconstructions. Comparison of volume reconstructions after 2D segmentation using: Baseline U-Net, and Edge2Prompt (trained only on MR T1-DUAL OOP). Predicted volumes are overlaid on the green ground truth in frontal and axial views [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 25 canonical work pages

[1]

Recent advances in computerized imaging and its vital roles in liver disease diagnosis, preoperative planning, and interventional liver surgery: A review,

P. Horkaew, J. Chansangrat, N. Keeratibharat, and D. C. Le, “Recent advances in computerized imaging and its vital roles in liver disease diagnosis, preoperative planning, and interventional liver surgery: A review,”World Journal of Gastroin- testinal Surgery, vol. 15, no. 11, p. 2382, 2023

work page 2023
[2]

Trends in medical imaging: from 2d to 3d,

G. Sakas, “Trends in medical imaging: from 2d to 3d,”Computers & Graphics, vol. 26, no. 4, pp. 577–587, 2002

work page 2002
[3]

Mri, ct scan, and ultrasound in the diagnosis of nonalcoholic fatty liver disease,

M. Noureddin, C. Khoyilar, and S. L. Palmer, “Mri, ct scan, and ultrasound in the diagnosis of nonalcoholic fatty liver disease,”Journal of clinical gastroenterology, vol. 49, no. 4, pp. 351–352, 2015

work page 2015
[4]

Accessible magnetic resonance imaging: a review,

S. Geethanath and J. T. Vaughan Jr, “Accessible magnetic resonance imaging: a review,”Journal of Magnetic Resonance Imaging, vol. 49, no. 7, pp. e65–e77, 2019

work page 2019
[5]

Various image segmentation techniques: a review,

D. Kaur and Y. Kaur, “Various image segmentation techniques: a review,”In- ternational Journal of Computer Science and Mobile Computing, vol. 3, no. 5, pp. 809–814, 2014

work page 2014
[6]

Survey of image edge detection,

R. Sun, T. Lei, Q. Chen, Z. Wang, X. Du, W. Zhao, and A. K. Nandi, “Survey of image edge detection,”Front. Signal Process., vol. 2, Mar. 2022

work page 2022
[7]

A comprehensive survey of multi-level thresholding segmentation methods for image processing,

M. Amiriebrahimabadi, Z. Rouhi, and N. Mansouri, “A comprehensive survey of multi-level thresholding segmentation methods for image processing,”Arch. Com- put. Methods Eng., vol. 31, pp. 3647–3697, Aug. 2024

work page 2024
[8]

Survey of clustering techniques enhancing image segmentation process,

S. Gangwar and R. P. Chauhan, “Survey of clustering techniques enhancing image segmentation process,” in2015 Second International Conference on Advances in Computing and Communication Engineering, pp. 34–39, IEEE, May 2015

work page 2015

Show all 29 references

[9]

Advances in medical im- age segmentation: a comprehensive review of traditional, deep learning and hybrid approaches,

Y. Xu, R. Quan, W. Xu, Y. Huang, X. Chen, and F. Liu, “Advances in medical im- age segmentation: a comprehensive review of traditional, deep learning and hybrid approaches,”Bioengineering, vol. 11, no. 10, p. 1034, 2024

2024
[10]

Chapter 13 - medical image segmentation using artificial intelligence,

A. Subasi, “Chapter 13 - medical image segmentation using artificial intelligence,” inApplications of Artificial Intelligence in Healthcare and Biomedicine(A. Subasi, ed.), Artificial Intelligence Applications in Healthcare and Medicine, pp. 377–400, Academic Press, 2024

2024
[11]

Intraoperative CT augmentation forneedle-basedliverinterventions,

S. El hadramy, J. Verde, N. Padoy, and S. Cotin, “Intraoperative CT augmentation forneedle-basedliverinterventions,” inLecture Notes in Computer Science,Lecture notes in computer science, pp. 291–301, Cham: Springer Nature Switzerland, 2023

2023
[12]

U-net: Convolutional networks for biomedical image segmentation,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,”CoRR, vol. abs/1505.04597, 2015

2015 arXiv
[13]

The domain shift problem of medical image segmentation and vendor-adaptation by unet-gan,

W. Yan, Y. Wang, S. Gu, L. Huang, F. Yan, L. Xia, and Q. Tao, “The domain shift problem of medical image segmentation and vendor-adaptation by unet-gan,” inMedical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, O...

2019
[14]

Foundation models for biomedical image segmentation: A survey,

H. H. Lee, Y. Gu, T. Zhao, Y. Xu, J. Yang, N. Usuyama, C. Wong, M. Wei, B. A. Landman, Y. Huo,et al., “Foundation models for biomedical image segmentation: A survey,”arXiv preprint arXiv:2401.07654, 2024. 10 Hollet et al

2024 arXiv
[15]

Foundational models in medical imaging: A comprehensive survey and future vision,

B. Azad, R. Azad, S. Eskandari, A. Bozorgpour, A. Kazerouni, I. Rekik, and D. Merhof, “Foundational models in medical imaging: A comprehensive survey and future vision,” 2023

2023
[16]

Sam 2: Segment anything in images and videos,

N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V. Alwala, N. Carion, C.-Y. Wu, R. Girshick, P. Dollár, and C. Feichtenhofer, “Sam 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024

2024 arXiv
[17]

Segment anything model for medical image seg- mentation: Current applications and future directions,

Y. Zhang, Z. Shen, and R. Jiao, “Segment anything model for medical image seg- mentation: Current applications and future directions,”Computers in Biology and Medicine, p. 108238, 2024

2024
[18]

Segment anything model for medical image analysis: an experimental study,

M. A. Mazurowski, H. Dong, H. Gu, J. Yang, N. Konz, and Y. Zhang, “Segment anything model for medical image analysis: an experimental study,”Medical Image Analysis, vol. 89, p. 102918, 2023

2023
[19]

Segment anything in medical images,

J. Ma, Y. He, F. Li, L. Han, C. You, and B. Wang, “Segment anything in medical images,”Nature Communications, vol. 15, no. 1, p. 654, 2024

2024
[20]

How to build the best medical image segmentation algorithm using foundation models: a comprehensive empirical study with segment anything model,

H. Gu, H. Dong, J. Yang, and M. A. Mazurowski, “How to build the best medical image segmentation algorithm using foundation models: a comprehensive empirical study with segment anything model,”Machine Learning for Biomedical Imaging, vol. 3, p. 88–120, May 2025

2025
[21]

Medical sam adapter: Adapting segment anything model for medical image segmentation,

J. Wu, Z. Wang, M. Hong, W. Ji, H. Fu, Y. Xu, M. Xu, and Y. Jin, “Medical sam adapter: Adapting segment anything model for medical image segmentation,” Medical image analysis, vol. 102, p. 103547, 2025

2025
[22]

Self-prompting large vision models for few-shot medical image segmentation,

Q. Wu, Y. Zhang, and M. Elbatel, “Self-prompting large vision models for few-shot medical image segmentation,” inMICCAI Workshop on Domain Adaptation and Representation Transfer, pp. 156–167, Springer, 2023

2023
[23]

Focal loss for dense object detection,

T. Lin, P. Goyal, R. B. Girshick, K. He, and P. Dollár, “Focal loss for dense object detection,”CoRR, vol. abs/1708.02002, 2017

2017 arXiv
[24]

A survey of loss functions for semantic segmentation,

S. Jadon, “A survey of loss functions for semantic segmentation,” in2020 IEEE Conference on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB), p. 1–7, IEEE, Oct. 2020

2020
[25]

Super-resolution in medical imaging,

H. Greenspan, “Super-resolution in medical imaging,”The computer journal, vol. 52, no. 1, pp. 43–63, 2009

2009
[26]

Super resolution techniques for medical image pro- cessing,

J. S. Isaac and R. Kulkarni, “Super resolution techniques for medical image pro- cessing,” in2015 International Conference on Technologies for Sustainable Devel- opment (ICTSD), pp. 1–6, IEEE, 2015

2015
[27]

CHAOS - Com- bined (CT-MR) Healthy Abdominal Organ Segmentation Challenge Data,

A. E. Kavur, M. A. Selver, O. Dicle, M. Barış, and N. S. Gezer, “CHAOS - Com- bined (CT-MR) Healthy Abdominal Organ Segmentation Challenge Data,” Apr. 2019

2019
[28]

Kornia: an open source differentiable computer vision library for pytorch,

D. M. E. Riba, D. Ponsa, E. Rublee, and G. Bradski, “Kornia: an open source differentiable computer vision library for pytorch,” inWinter Conference on Ap- plications of Computer Vision, 2020

2020
[29]

A descriptive algorithm for sobel image edge detection,

O. R. Vincent, O. Folorunso,et al., “A descriptive algorithm for sobel image edge detection,” inProceedings of informing science & IT education conference (In- SITE), vol. 40, pp. 97–107, 2009

2009

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Recent advances in computerized imaging and its vital roles in liver disease diagnosis, preoperative planning, and interventional liver surgery: A review,

P. Horkaew, J. Chansangrat, N. Keeratibharat, and D. C. Le, “Recent advances in computerized imaging and its vital roles in liver disease diagnosis, preoperative planning, and interventional liver surgery: A review,”World Journal of Gastroin- testinal Surgery, vol. 15, no. 11, p. 2382, 2023

work page 2023

[2] [2]

Trends in medical imaging: from 2d to 3d,

G. Sakas, “Trends in medical imaging: from 2d to 3d,”Computers & Graphics, vol. 26, no. 4, pp. 577–587, 2002

work page 2002

[3] [3]

Mri, ct scan, and ultrasound in the diagnosis of nonalcoholic fatty liver disease,

M. Noureddin, C. Khoyilar, and S. L. Palmer, “Mri, ct scan, and ultrasound in the diagnosis of nonalcoholic fatty liver disease,”Journal of clinical gastroenterology, vol. 49, no. 4, pp. 351–352, 2015

work page 2015

[4] [4]

Accessible magnetic resonance imaging: a review,

S. Geethanath and J. T. Vaughan Jr, “Accessible magnetic resonance imaging: a review,”Journal of Magnetic Resonance Imaging, vol. 49, no. 7, pp. e65–e77, 2019

work page 2019

[5] [5]

Various image segmentation techniques: a review,

D. Kaur and Y. Kaur, “Various image segmentation techniques: a review,”In- ternational Journal of Computer Science and Mobile Computing, vol. 3, no. 5, pp. 809–814, 2014

work page 2014

[6] [6]

Survey of image edge detection,

R. Sun, T. Lei, Q. Chen, Z. Wang, X. Du, W. Zhao, and A. K. Nandi, “Survey of image edge detection,”Front. Signal Process., vol. 2, Mar. 2022

work page 2022

[7] [7]

A comprehensive survey of multi-level thresholding segmentation methods for image processing,

M. Amiriebrahimabadi, Z. Rouhi, and N. Mansouri, “A comprehensive survey of multi-level thresholding segmentation methods for image processing,”Arch. Com- put. Methods Eng., vol. 31, pp. 3647–3697, Aug. 2024

work page 2024

[8] [8]

Survey of clustering techniques enhancing image segmentation process,

S. Gangwar and R. P. Chauhan, “Survey of clustering techniques enhancing image segmentation process,” in2015 Second International Conference on Advances in Computing and Communication Engineering, pp. 34–39, IEEE, May 2015

work page 2015